Parallelization

[salleuska]

Suggestion from Perry for parallelization

Build a small collection of compiled nimble objects (model, MCMC, calculator,
simulate function), one per core, and let each core keep its own across one
replicate after another. For example, for 100 replicates on 5 cores, that is 5
compilations, not 100. Compiled nimble objects cannot be sent between
processes, so each core builds its own; the point is that it builds it once.

Suggested tool: the future package.

To do:

• Make the collection buildable from the specifications alone, so a core can
  build its own with nothing passed in.
• Add future to Suggests, and keep the serial loop working without it.
• Decide what gets reset between replicates.